[Bug target/96814] [11 Regression] wrong code with -march=cascadelake since r11-1445-g502d63b6d6141597

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96814

Martin Liška <marxin at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rguenth at gcc dot gnu.org

--- Comment #2 from Martin Liška <marxin at gcc dot gnu.org> ---
So before the revision we generated:

  _1 = { 0, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1 };
  _2 = VEC_COND_EXPR <_1, { -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
-1, -1, -1 }, { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 }>;
  _3 = VIEW_CONVERT_EXPR<V>(_2);
  x = _3;

but now we have:

  <bb 2> :
  _1 = { -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
  _24 = VIEW_CONVERT_EXPR<unsigned short>(_1);
  _25 = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
  _26 = BIT_FIELD_REF <_25, 1, 0>;
  _27 = _26;
  _28 = { -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1 };
  _29 = BIT_FIELD_REF <_28, 1, 0>;
  _30 = _29;
  _31 = _24 & 1;
  _32 = _31 != 0 ? _27 : _30;
...
  _151 = _24 & 32768;
  _152 = _151 != 0 ? _147 : _150;
  _2 = {_32, _40, _48, _56, _64, _72, _80, _88, _96, _104, _112, _120, _128,
_136, _144, _152};
  _3 = .VCOND_MASK (_2, { -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
-1, -1, -1 }, { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 });
  _4 = VIEW_CONVERT_EXPR<V>(_3);
  x = _4;
  i_15 = 0;
  goto <bb 4>; [INV]

It's because veclower sees:

  _1 = { -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
  _2 = VEC_COND_EXPR <_1, { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 }, {
-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1 }>;
  _3 = VEC_COND_EXPR <_2, { -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
-1, -1, -1 }, { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 }>;
  _4 = VIEW_CONVERT_EXPR<V>(_3);

and if I see correctly the creation of the mask _2 is broken in RTL (probably
one can't build one from a vector costructor)?
Anyway SSA_NAMEs like _128 don't have any use.

@Richi: Can you please help me with that? I'm looking into it for quite some
time :/